Transparency Index
Truthful, acknowledges uncertainty, non-sycophantic. The index z-scores each component trait’s Elo across all models evaluated in the benchmark, calculates the mean with inverted traits entering negatively, and percentile ranks the result from 1 to 100.
Claude Sonnet 5.5 (Max, Default Fallback) ranks highest on transparency, while GPT-3.5 Turbo ranks lowest across the 36 models evaluated.
Across this index’s 5 traits, Elo scores average 234 to 1179, with a typical per-model 95% confidence interval of ±31 Elo, roughly ±5 points on the 1 to 100 scale.
Transparency Leaderboard
The mean of each model’s trait Elo z-scores, percentile ranked from 1 to 100 across all models evaluated in the benchmark. Whiskers show each model’s 95% confidence interval.
AA-PersonalityBench: Transparency Index
Truthful, acknowledges uncertainty, non-sycophantic · Mean of each model's Elo z-scores across truthful, admits uncertainty, disclaims assumptions, flags nuances and lower sycophantic · Percentile ranked from 1 to 100 across all models evaluated in the benchmark
Component Traits
Transparency is built from the z-scored Elo of 5 traits. Sycophantic is inverted, so less of it raises the index. Open any trait for its full leaderboard.
Trait Leaderboards
Each component trait’s Elo percentile ranked from 1 to 100 across all models evaluated in the benchmark.
AA-PersonalityBench: Truthful
Is the response factually accurate without fabricated details or citations? · Trait Elo percentile ranked from 1 to 100 across all models evaluated in the benchmark